Should you use AI-generated characters for close-up acting shots?

No — keep AI-generated characters out of close-up acting shots. AI holds up on wide environment shots, but the moment you push in on a character performing, faces and micro-expressions read as fake. Pull out rather than push in: shoot your actor practically for acting beats and let AI build the wide world around them.

Reserve AI generation for wide environment shots and keep any real acting on a practically-shot actor. Filmmaker Alex Arfaoui, who mixed AI and practical footage in a short film, states the boundary directly: "If you push into this shot, or if the shot is too close, and you're going to get the actual character doing any kind of acting, that's when things look fake." The working rule is pull out, never push in — the wider the frame, the less the AI has to render believable facial performance, and the more convincing the shot.

Build the world outward from your practical shot instead. Lock the close-up you filmed yourself as the anchor, then generate wide shots that place the same action in a larger environment. In one documented production, an emergence-from-water shot was filmed in a 6x6 ft studio pool — too small to hide the set edges, and the actor could tread water for only one to two minutes per take — so the filmmaker shot the tight practical takes and used the invideo agent (running Seedance 2.0) to generate the wide red-ocean and shoreline shots around them. The audience sees a real face acting in close-up and an AI-generated world in the wides, and the seam disappears.

Let story logic pick which wide shots to generate. Choose AI shots for narrative continuity, not visual impressiveness — if your practical close-up shows a character crawling out of water, a shoreline wide matches the action and cuts cleanly. Carry environmental continuity cues (matching haze, light direction, water color) between the practical and generated shots so transitions feel like one location.

If a generated shot must include your character, anchor it to your own footage. Upload screenshots and stills from your practical shoot into the invideo agent as visual references before generating — reference-anchored output stays close to your real color, contrast, and look, where text-only prompts drift. When a generation lands close, pin it and request variations: "create another one like it, but change X, Y, and Z." Keep the character small in frame and non-performing.

Blend everything in post. Color grade AI shots to match your practical footage and layer sound design over them — sound is the multiplier that makes generated inserts feel physical and real. Footage that still reads as artificial at normal speed can be salvaged as fast cutaway inserts by raising playback speed 1,000–5,000% and splicing the fragments, but that trick works for energy inserts, not acting shots — for performance, the answer stays practical.

Watch some of these to see what works for you:

If you push into this shot, or if the shot is too close, and you're going to get the actual character doing any kind of acting, that's when things look fake.

— Alex Arfaoui, filmmaker